Meta's 2026 AI Transformation: Planned Team Cuts Up to 60%, Code Output Up 220%, But Productivity Gains Lag｜Article

Reuters reveals Meta's internal AI transformation under Project OT, which explored cutting some teams by up to 60% while increasing AI-generated code by 220%. However, actual user-facing features only grew 36%, with rising incidents and firefighting time. Meta halted a second round of layoffs…

Tonight's news is not about predicting:

"AI might one day replace jobs."

It's more direct.

A tech company that's among the richest, most AI-savvy, and actively building AI agents itself,

is truly redesigning its organization according to the idea:

"AI can do more work for people."

The result?

The first round already ran into reality.

The company is:

Meta.

Meta's original plan was more than just adopting chatbots for employees

Reuters today published an extensive investigation that unveiled an internal Meta initiative codenamed:

Project OT

OT stands for:

Organization Transformation.

Simply put:

Meta wanted to redesign exactly what a company should look like in the AI era.

Not just giving 10 people an AI tool each,

but asking directly:

If AI Agents could handle a large portion of routine work,

does the company even need as many people as before?

The most radical scenario: Some teams cut by up to 60%

According to internal documents and sources seen by Reuters,

Project OT's planning considered scenarios where:

some teams were reduced by as much as 60%.

Note:

This does not mean Meta was planning to cut 60% of its entire workforce.

Meta itself clarified:

The most aggressive 60% reduction scenario involved only:

certain teams.

And this involved a mix of:

layoffs, hiring freezes,

and redeploying employees.

It is incorrect to report:

"Meta to cut 60% of staff."

But the overall direction was indeed very aggressive

Meta envisioned transforming traditional product development teams,

which might previously require:

10 to 20 people,

into much smaller:

3–5 person pods.

Roles would blend, such as:

engineers, product managers, designers, data scientists,

without the usual layered hierarchical workflows.

Small teams working with:

AI tools,

AI Agents,

and shared experts,

aiming to handle:

more work with fewer people.

This concept represents an "AI-native company"

Meaning not simply:

the company has purchased ChatGPT,

but the whole work design assumes from the start:

AI is part of the workforce.

For example, AI:

conducts analysis,

writes code,

helps prioritize,

builds prototypes,

and even works alongside a few humans as virtual workers.

If successful,

then naturally, the old org chart might be too big.

Project OT had planned two waves of restructuring

The internal plans Reuters reviewed showed:

The first wave was scheduled for May,

the second for November.

On May 20,

Meta proceeded with roughly:

10% workforce reduction.

But one thing suddenly changed.

Just hours before the first layoffs, after discussions with Zuckerberg and senior leadership,

the planned second wave company-wide restructuring was canceled.

Why the sudden stop?

Reuters states it is unclear which exact factor led Zuckerberg to change course.

So it’s not accurate to say:

"AI failed so Meta stopped layoffs."

But several issues arose simultaneously.

First problem: Strong employee backlash

Meta employees gradually realized the company was saying:

AI makes work more efficient,

while simultaneously implementing:

layoffs, team downsizing, position redesign,

and even collecting employee mouse clicks and keystrokes

to train AI on how to operate computers.

Many developed a direct thought:

"Am I training an AI to replace myself?"

Employee morale also started dropping at Meta

Internal Pulse Surveys obtained by Reuters showed positive employee sentiment declined from:

74%

to:

55%

This isn’t a minor shift.

For a huge company-wide transformation requiring full cooperation,

employees doubting management’s intentions

posed a critical risk.

But tonight, morale isn’t the most important issue

The really significant numbers are:

220% vs 36%

After intensively using AI to generate code internally,

Meta’s internal data shows software platform and infrastructure code changes increased by:

220% year over year.

By this metric alone, AI looks like a massive success.

Engineering activity surged.

However, user-facing features only increased by 36%

The actual changes delivered to Meta users—in the form of new or upgraded features—

only grew by:

36%

Comparing these two figures immediately highlights the issue.

More code changes do not equal a proportional increase in product value.

This is the first common pitfall of AI productivity

Confusing:

Activity

with:

Outcome.

For example, if an engineer used to write 100 lines of code a day,

and now AI helps write 1,000 lines,

it’s easy to say productivity increased tenfold.

But customers don’t care about how many lines are written.

They care whether features are launched:

faster, more stable, easier to use, and solve real problems.

Code itself is not the product

Like a factory using three times as many screws per day cannot prove triple car output,

the most dangerous AI Coding KPI is:

how much code, pull requests, commits, or prototypes the model generates.

These are all intermediate outputs.

The real metrics should be:

features shipped,

quality,

failure rate,

customer value,

development cycles,

and cost.

Meta also faced a second problem: AI-generated code triggered reliability warnings

Meta internal documents obtained by Reuters indicate as early as March,

the infrastructure team warned about:

an AI Coding Surge

triggering:

Reliability Warning Signs.

By April, additional warnings noted:

unrestricted AI Agents started performing:

disruptive operations that human engineers normally wouldn’t execute in bulk at once.

The subsequent numbers are even more telling

Internal Meta data cited by Reuters showed:

major technology and security incidents increased about:

40% year over year,

while engineering time spent firefighting—

handling incidents and restoring systems—

rose by:

70%.

Meta declined to comment on these internal incident figures,

so these are numbers Reuters acquired internally,

not official Meta KPIs.

This reveals the second illusion about AI Coding

AI speeding up first drafts does not mean the entire workflow speeds up.

If AI takes 10 minutes to write code,

but engineers spend 2 hours debugging, reviewing, rolling back, and handling incidents,

the initial 10 minutes can’t be the only metric tracked.

The real cost must include firefighting time

This matches earlier AI ROI discussions.

If AI output only costs $0.02 per run,

it looks cheap,

but if 10% of results require human rework,

5% cause errors,

and 1% lead to formal incidents,

the actual cost isn’t just token fees.

Third problem: AI Agents haven’t matured as fast as management hoped

In July at a Meta internal Town Hall meeting,

Zuckerberg acknowledged a key point:

AI Agent technology’s acceleration pace was not as fast as originally expected.

He still believed improvements would come in the coming months.

But this meant a core assumption behind the organizational overhaul—

"Agents can quickly take over a large volume of tasks."

was overly optimistic.

This is the third major risk of enterprise AI transformation

Confusing:

future capabilities

with:

current capabilities.

For example, watching a demo that AI can:

write code, conduct research, make presentations, operate tools,

might lead management to calculate:

one person with AI equals three people,

so teams can be cut by two-thirds.

The problem is a demo proves:

something can be done,

not that it can be done reliably every day in production environments handling all exceptions.

This echoes yesterday’s Physics AI deployment issue

A model performing well in test environments

is different from operating robustly in production daily.

Companies need to manage:

anomalous data, system failures, permissions, security, edge cases, team handoffs, human judgments—

all happening simultaneously.

Meta is actually one of the best companies in the world to test this

Making tonight’s news even more noteworthy.

Meta doesn’t lack:

AI models,

engineers,

GPUs,

data,

AI researchers,

or funding.

They plan to invest hundreds of billions on AI infrastructure this year.

If even Meta finds:

the leap from "AI can write lots of code" to "the company really needs fewer people"

is a huge gap,

then ordinary companies shouldn’t make layoff decisions solely based on demos.

This does not mean Meta is giving up on AI-native

And don’t interpret tonight’s news as:

"Meta admits AI is useless."

Not at all.

Meta continues to heavily invest in AI,

reorganize teams,

deploy employees to more important new roles,

and form smaller, agile development pods.

Some internal data production teams already support new model training.

The direction remains;

only speed and aggressiveness have changed.

Meta hasn’t abandoned the long-term vision of "small teams + AI"

When Zuckerberg spoke publicly about AI’s future recently,

he still predicted AI could make companies smaller,

with just a few people running companies that previously required many employees.

However, his argument is:

average company size shrinking doesn’t necessarily mean:

overall job opportunities decrease,

since AI may enable:

more new companies,

more new services,

and more new occupations.

That’s why asking if "AI causes layoffs" is too simple

What may actually happen is:

Type 1 companies: no layoffs, AI used to do more work.

Type 2: smaller headcount, same revenue.

Type 3: redeploy staff to new roles.

Type 4: some jobs disappear, new ones created.

Type 5: AI efficiency underperforms, eventually rehiring needed.

The world won’t be just:

"replacement"

"no replacement."

Meta’s experiment is really testing a major question

How fast can a company shrink?

Technology can improve tenfold in a year,

but an organization can’t just hit Update and finish,

because organizations involve:

skills,

trust,

responsibility,

experience,

communication,

culture,

and domain knowledge.

None of these automatically upgrade just because model benchmarks improve.

The small pod idea itself is not wrong

Three or four experts,

working with powerful AI,

can indeed accomplish work that used to require a dozen people.

Many startups prove this.

But:

Designing from scratch as a startup

is very different from

slicing a large company of tens of thousands into startup-style pods.

Because big companies have a lot of invisible coordination work

Who approves features?

Who handles security?

Who checks privacy?

Who manages infrastructure?

Who knows why a certain piece of code from five years ago can’t be changed?

Who understands local market regulations?

Who knows changing a feature will impact another product?

Most of this isn’t visible from a GitHub code snippet.

AI most easily replaces "steps in a workflow"

But companies function via a network of responsibilities.

For example:

AI can write login functionality.

But if login fails at midnight,

who is responsible then?

AI can update recommendation models.

But if this causes大量錯誤內容 (many errors),

who decides to roll back?

AI may perform product analysis.

But if these conclusions impact multi-million-dollar investments,

who signs off?

Work can be automated,

but responsibility cannot just vanish.

This explains why "AI-native" doesn’t just mean "fewer people is better"

A truly mature AI-native company should ask:

Which tasks is AI clearly faster at?

Hand them over to AI.

Which tasks can AI do but are error-prone and easy to check?

AI does first; humans verify.

Which tasks have costly consequences if wrong?

Increase testing, approval, and permissions.

Which tasks require cross-department responsibility?

Don't remove humans just because AI can do the steps.

This is the real way to redesign work.

The 220% and 36% figures should be posted on every company’s wall

After AI rollout,

if your dashboard only shows:

number of generated articles,

code produced,

prompts completed,

agent runs,

or user numbers,

you might prematurely declare:

AI implementation a success.

But a second dashboard should ask:

how much feature delivery improved?

how much customer problems decreased?

how much revenue increased?

how much cycle time reduced?

did errors increase or decrease?

how much rework and incidents happened?

This is the true meaning of productivity

It’s not:

doing more.

It’s:

producing more genuinely valuable outcomes with the same resources.

If:

output volume +220%,

valuable outcomes +36%,

incidents +40%,

firefighting time +70%,

then it’s time to ask:

Are we really faster, or just producing more to fix?

This question is even more important than layoffs themselves

If AI were reliable, stable, low-cost, and highly productive,

companies would naturally recalculate their staffing needs.

That’s normal.

But if productivity is not accurately measured,

and a false assumption like "one person + AI equals three people" is made,

then cutting staff risks losing team experience that’s hard to recover quickly.

Especially don’t prove engineer cuts by just more code output

Because engineers do much more than type, including:

understanding problems,

designing architecture,

reviewing,

debugging,

security,

maintenance,

communication,

incident handling,

and deciding what not to do.

AI speeding up typing is important,

but it cannot justify proportional elimination of entire roles.

Employee backlash also highlights a key AI adoption issue: trust

If a company asks employees:

Please document your entire workflow for AI to learn from,

but employees think:

Will learning this mean I get laid off?

What will happen?

Not cooperative AI transformation.

Employees may:

resist,

withhold knowledge,

avoid testing,

even intentionally avoid using AI.

In the end, the company won’t get the critical domain knowledge it needs.

So AI transformation requires a new social contract

Companies can’t just say:

"AI makes you more efficient."

Employees will then ask:

"What do I get from this increased efficiency?"

More work?

New skills?

Promotions?

Higher pay?

Less repetitive tasks?

Or layoffs?

This isn’t an HR or PR problem,

but integral to AI adoption itself.

SasaDaily previously discussed another Meta CTO remark

Andrew Bosworth said he hopes that AI productivity gains will lead to:

more products,

not just:

more vacation time.

His article asked:

Who ultimately benefits from the time AI saves?

Tonight's Reuters investigation

pushes this question deeper.

If AI convinces the company people can be fewer, then:

how to measure productivity?

When can headcount truly shrink?

If measured wrong,

who pays for rework and incidents?

Japanese companies’ AI adoption offers another lesson

Many firms have:

"someone using AI"

versus:

"the entire company truly transformed by AI,"

which are very different.

Meta now seems to have come to the other extreme:

not stopping at employees using chatbots,

but directly attempting:

organizational redesign.

This proves the hardest phase of enterprise AI:

is never buying tools,

but changing processes,

responsibilities,

org charts,

and KPIs.

What can typical companies learn tonight?

Not that:

Meta made a mistake, so don’t adopt AI.

But rather, the adoption sequence:

mustn’t be reversed.

Stage 1: Test work tasks

For example:

Can AI reduce customer service documentation time by 50%?

Can AI Coding shorten bug fix time?

Can AI Agents complete fixed processes?

Stage 2: Measure total costs

Don’t only measure:

how long AI runs,

but also review human checking,

modifications,

failures,

rollbacks,

and incidents.

Stage 3: Measure actual outcomes

Examples:

feature delivery speed,

customer problem resolution,

revenue,

cost,

quality.

Stage 4: Then redesign organization

If after half a year of data shows some work can reliably be done with half the staff,

then consider reorganizing,

not necessarily cutting half the team,

but redeploying people to tasks the company couldn’t support before.

Because AI’s most valuable results aren’t always fewer people

Could be that a month’s feature output triples,

customer service expands from only Mandarin to 10 languages,

or small companies gain analytics capabilities they lacked.

These are all AI ROI.

The most important takeaway isn’t how many layoffs Meta almost made

But that a global leader heavily investing in AI encountered:

massive AI activity growth does not equal proportional productivity growth.

This is a problem all enterprises face now.

The most dangerous new AI KPI may be "looking busy"

Agent runs: 100,000 times.

Code up 220%.

Documents up 500%.

Content up 1,000%.

If customer value only increases slightly,

while incidents, rework, and management burdens rise,

AI is just making the company:

produce more activity faster.

A truly mature AI-native company should do the opposite

Not ask:

"Can AI let me have 60% fewer people?"

But first ask:

"Which outcomes can sustainably be delivered reliably with fewer people?"

Mixing up this order changes the whole risk profile.

One concluding thought tonight

AI will very likely:

make many companies smaller,

change many roles,

and eliminate many workflow steps.

This is inevitable.

But Meta's internal experiment reminds us:

generating more output with AI

is still quite far from:

reliably replacing entire human jobs.

True enterprise AI transformation

starts not by calculating who to cut,

but by proving:

outputs increased,

quality maintained,

incidents decreased,

and firefighting time truly reduced.

When these hold, organizations will naturally evolve.

If not,

no flashy AI-native org chart is more than a hypothesis.

Today, progress together with AI.

Learn an AI skill daily.

Save a bit of time daily.

Improve a bit every day.

SasaDaily, growing with you.